US 11,790,533 B2
Machine learning based image segmentation training with contour accuracy evaluation
Mengyu Chen, Santa Barbara, CA (US); Miaoqi Zhu, Studio City, CA (US); Yoshikazu Takashima, Los Angeles, CA (US); Ouyang Chao, Torrance, CA (US); Daniel De La Rosa, Los Angeles, CA (US); Michael Lafuente, Los Angeles, CA (US); and Stephen Shapiro, Torrance, CA (US)
Assigned to Sony Group Corporation, Tokyo (JP); and Sony Pictures Entertainment Inc., Culver City, CA (US)
Filed by Sony Group Corporation, Tokyo (JP); and Sony Pictures Entertainment Inc., Culver City, CA (US)
Filed on Feb. 18, 2021, as Appl. No. 17/179,061.
Claims priority of provisional application 63/047,750, filed on Jul. 2, 2020.
Prior Publication US 2022/0005200 A1, Jan. 6, 2022
Int. Cl. G06T 7/12 (2017.01); G06V 10/44 (2022.01); G06F 18/214 (2023.01); G06V 10/774 (2022.01)
CPC G06T 7/12 (2017.01) [G06F 18/214 (2023.01); G06V 10/44 (2022.01); G06V 10/774 (2022.01); G06T 2207/10024 (2013.01)] 15 Claims
OG exemplary drawing
 
1. A method for improving the accuracy of predicted segmentation masks, the method comprising:
extracting a ground-truth red-green-blue (RGB) image buffer and a binary contour image buffer from a ground-truth RGB image container for segmentation training;
generating predicted segmentation masks from the ground-truth RGB image buffer;
generating second binary contours from the predicted segmentation masks using a particular algorithm;
computing a segmentation loss between manually-segmented masks of the ground-truth RGB image buffer and the predicted segmentation masks;
computing a contour accuracy loss between contours of the binary contour image buffer and the second binary contours of the predicted segmentation masks;
computing a total loss as a weighted average of the segmentation loss and the contour accuracy loss; and
generating improved binary contours by compensating the contours of the binary contour image buffer with the computed total loss, wherein the improved binary contours are used to improve the accuracy of the predicted segmentation masks.